AWhat changed
EAGLE conditioned feature-level drafts on one-step-advanced tokens; EAGLE2 selected context-dependent draft trees from confidence scores; EAGLE3 removed feature-regression constraints and trained multi-layer-fused token drafts using training-time rollout tests.
Scientific scope
EAGLE conditioned feature-level drafts on one-step-advanced tokens; EAGLE2 selected context-dependent draft trees from confidence scores; EAGLE3 removed feature-regression constraints and trained multi-layer-fused token drafts using training-time rollout tests. One family with three source-specific deltas, not three inventions of speculative sampling. Losslessness depends on the inherited exact verification/sampling procedure and implementation; reported speedups are model, workload, hardware and batch dependent. EAGLE2 avoids additional training beyond EAGLE, not all draft-model training; EAGLE3 changes and trains the drafter.
Evidence
The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. One family with three source-specific deltas, not three inventions of speculative sampling. Losslessness depends on the inherited exact verification/sampling procedure and implementation; reported speedups are model, workload, hardware and batch dependent. EAGLE2 avoids additional training beyond EAGLE, not all draft-model training; EAGLE3 changes and trains the drafter.
Contributors
- Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.
- EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty
- EAGLE-2: Faster Inference of Language Models with Dynamic Draft Trees
- EAGLE-3: Scaling up Inference Acceleration of Large Language Models via Training-Time Test
Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.
Full scientific record · Browse contribution grades · Request a correction